Gemma 4 by Google: how to run local AI on your phone or computer in Ukrainian

What is Gemma 4, and how is it different from Gemini?

Gemma 4 is Google DeepMind’s family of open-weight models. You can download and run them on your own device. Gemini is a separate line of Google products and models; local Gemma is not an offline copy of cloud Gemini. This guide was checked against documentation on October 28, 2026.

Gemma 4 handles text and images; E2B, E4B and 12B also support audio input at the model level. Available features depend on the app and model file. It generates text, not a video or image file. It can help with ideas, writing, explanations and code, but can invent facts.

Does Gemma understand Ukrainian?

Google describes Gemma 4 as a multilingual family pretrained on over 140 languages. Ukrainian evaluations include larger Gemma 4 models, but that does not establish the same quality for mobile E2B or E4B. Small models may mix languages, use awkward phrasing or make mistakes. We have not tested them on your phone and cannot promise flawless Ukrainian.

Start with: “Answer in natural Ukrainian and do not switch to Russian. If you do not know, say so.” Ask it to summarize a Ukrainian text, explain a household manual and write a short letter. Compare the results: choosing an answer language does not change the model itself.

The simplest phone option: Google AI Edge Gallery

This is Google’s separate free experimental app for Android 12+ and iPhones running iOS 17+. It runs compatible models on the device, not in cloud Gemini. Installation and model downloads need internet; afterward, test ordinary chat in airplane mode. Some extra features, including online search, may need a connection.

1. Install Google AI Edge Gallery from the official Google Play or App Store links at the end of this article. Connect to Wi-Fi, charge your phone and free several gigabytes of storage.

2. Open AI Chat. In the list, choose Gemma-4-E2B-it, expand it and tap Download. If prompted, sign in to Hugging Face and accept the model terms. When it finishes, tap Try It. Availability and button names can vary by app version and device.

3. Ask a question in Ukrainian and request a Ukrainian response. Test several replies, then turn on airplane mode and repeat. If the model is unavailable, slow or crashes, update the app, close other apps, restart the phone and try a shorter chat or a smaller compatible model.

In Google’s Android list, Gemma-4-E2B-it is approximately 2.59 GB with a minimum of 8 GB device RAM; E4B is approximately 3.66 GB with 12 GB RAM. These are specific configurations, not universal requirements for all files or iPhones. Storage and RAM are different. Even a compatible phone may be slow or warm up.

How to install Gemma on Windows, Mac or Linux

For a straightforward setup without commands, use LM Studio. Install it from the official site, open Discover and search for google/gemma-4-e2b. Download an available local version, then open Chat, select the model and load it into memory. Ask in Ukrainian. If prompted, download the runtime while online.

LM Studio lists E2B at about 4.2 GB, E4B at 5.9 GB and 12B at 7.4 GB. These are particular download sizes, not sufficient RAM: the system and context need memory too. Files and requirements may change. Start with E2B; if your computer handles it, compare a larger model’s answers.

If you are comfortable with a terminal, install Ollama and run ollama run gemma4:e2b. The first download requires internet. Choose the local tag without the :cloud suffix; after downloading, test it without a network.

What if the Ukrainian is poor or the app crashes?

Try MamayLM-Gemma-3-4B-IT: this is a separate adaptation of the older Gemma 3, specifically trained further for Ukrainian and English. It runs in compatible GGUF apps, including LM Studio on a computer or LM Playground on Android. The Q4_K_M file is about 2.49 GB but may still exceed the phone’s available RAM. It is not Gemma 4 and is not guaranteed to be more stable.

You cannot simply open MamayLM GGUF in Google AI Edge Gallery: it uses compatible LiteRT files, including .litertlm. If an alternative app also crashes, reduce context or try a lighter model; test its Ukrainian again.

What works offline and where to get knowledge

A downloaded local model does not automatically receive current news or the whole of Wikipedia. For a large offline library, install Kiwix separately and download Ukrainian Wikipedia and the reference material you need. Kiwix does not automatically connect to Gallery: open the source and copy a passage into chat if useful. For medicine, safety and survival, check primary guidance; the model can be wrong.

Sources and downloads

• Google AI Edge Gallery — Android
• Google AI Edge Gallery — iPhone
• Google AI Edge Gallery — instructions
• Google — Android models and memory
• Google — Gemma 4 documentation
• LM Studio — Gemma 4 models
• LM Studio — download
• Ollama — Gemma 4
• MamayLM — Ukrainian model
• UkrQualBench — Ukrainian language evaluation
• Offline library and Survival LLM